quantitative organizational modeling and design for multi-agent systems

QUANTITATIVE ORGANIZATIONAL MODELING AND DESIGN FOR MULTI-AGENT SYSTEMS FEBRUARY 2006 BRYAN HORLINGB.Sc., TRINITY COLLEGEM.S., UNIVERSITY OF MASSACHUSETTS AMHERSTPh.D., UNIVERSITY OF MAS

QUANTITATIVE ORGANIZATIONAL MODELING AND

DESIGN FOR MULTI-AGENT SYSTEMS

A Dissertation Presented

byBRYAN HORLING

Submitted to the Graduate School of theUniversity of Massachusetts Amherst in partial fulfillment

of the requirements for the degree ofDOCTOR OF PHILOSOPHY

February 2006Department of Computer Science

UMI Number: 3206188

3206188 2006

UMI Microform Copyright

All rights reserved This microform edition is protected against unauthorized copying under Title 17, United States Code.

ProQuest Information and Learning Company

300 North Zeeb Road P.O Box 1346 Ann Arbor, MI 48106-1346

by ProQuest Information and Learning Company

 Copyright by Bryan Horling 2006

All Rights Reserved

QUANTITATIVE ORGANIZATIONAL MODELING AND

DESIGN FOR MULTI-AGENT SYSTEMS

A Dissertation Presented

byBRYAN HORLING

Approved as to style and content by:

Victor Lesser, Chair

Brian Levine, Member

Shlomo Zilberstein, Member

Anna Nagurney, Member

W Bruce Croft, Department ChairDepartment of Computer Science

To my wife and partner, Maura.

I would like to thank the members of my committee, Brian Levine, Anna ney and Shlomo Zilberstein, for all their hard work Your questions helped guide

Nagur-me towards interesting and challenging paths, and our discussions helped fraNagur-me thebroader context of this work

My family has been my foundation throughout my graduate career, giving methe opportunity and encouragement I needed to see it through My wife and partnerMaura deserves the most praise, as the time I needed to do my work frequentlyrequired sacrifice on her part Your emotional support, patience, and willingness tolisten as I ramble on about my research were all invaluable I would like to thankMegan as well, for giving me the opportunity to get some writing done during yourMonday afternoon naps, for only throwing my papers on the floor once, and for beingthere to play with when I didn’t feel like working

My mother, father and sister have supported me for more than thirty years now,and I clearly would not be where I am today without their help There are too manypoints along the way that could be mentioned, but most of all it was your belief in

me that provided the motivation I needed to see me through

I would like to thank the Fennellys - Paul, Kathy, Jeff, Eileen and Kevin, for theirencouragement, and for providing a place of refuge for Maura and I when one or both

of us needed a break

Over the course of the last two years I had the opportunity to speak with manyindividuals about my work, and benefited greatly from these interactions In noparticular order I would like to thank Dan Corkill, Roger Mailler, Anita Raja, SheriefAbdallah, Mark Sims, Mark Fox, Neil Immerman, Les Gasser, Jim Kurose, HaizengZhang, Jiaying Shen, Raphen Becker, Katia Sycara, Frank Dignum, Virginia Dignum,Carl Hewitt and Tom Wagner for those conversations

I would like to thank Michele Roberts, without whom papers and forms wouldnot reach their destinations, salaries would not be paid, and life in the lab would

be generally less pleasant I would also like to thank the past and present bers of the Multi-Agent Systems Lab game night, which helped provide a neededdistraction almost every week: Sherief Abdallah, Peter Amstutz, Mike Atighetchi,Raphen Becker, Brett Benyo, Ross Fairgrieve, Roger Mailler, Stephen Murtagh, DanNeimann, Shichao Ou, Rodion Podorozhny, Kyle Rawlins, Jiaying Shen, Regis Vin-cent, Tom Wagner, and Ping Xuan I have also benefited from my interactions withthe many other members of the MAS lab, including Ana Bazzan, Andrew Fast, NadiaGhamrawi, AnYuan Guo, Frank Klassner, Hala Mostafa, Mike O’Neill, John Ostwald,Anita Raja, Zach Rubinstein, Haizheng Zhang, and Shelley Zhang

mem-I would like to single out Regis Vincent, Roger Mailler, Raphen Becker, JiayingShen and Kyle Rawlins for their help in designing and implementing the distributedsensor network platform analyzed in this work I would also like to thank HaizhengZhang for his research in information retrieval, which this work also exploits

The fine people at Freeverse Software (Ian, Colin & Steve) deserve thanks forgiving me a creative outlet Thanks also go to the Amherst College ComputingCenter and the authors of the JEP and JDOM software libraries

QUANTITATIVE ORGANIZATIONAL MODELING AND

DESIGN FOR MULTI-AGENT SYSTEMS

FEBRUARY 2006

BRYAN HORLINGB.Sc., TRINITY COLLEGEM.S., UNIVERSITY OF MASSACHUSETTS AMHERSTPh.D., UNIVERSITY OF MASSACHUSETTS AMHERST

Directed by: Professor Victor Lesser

As the scale and scope of distributed and multi-agent systems grow, it becomesincreasingly important to design and manage the participants’ interactions Thepotential for bottlenecks, intractably large sets of coordination partners, and sharedbounded resources can make individual and high-level goals difficult to achieve Toaddress these problems, many large systems employ an additional layer of structuring,known as an organizational design, that assigns agents particular and different roles,responsibilities and peers These additional constraints can allow agents to operateeffectively within a large-scale system, with little or no sacrifice in utility Differentdesigns applied to the same problem will have different performance characteristics,therefore it is important to understand and model the behavior of candidate designs

In the multi-agent systems community, relatively little attention has been paid tounderstanding and comparing organizations at a quantitative level In this thesis, I

show that it is possible to develop such an understanding, and in particular I show howquantitative information can form the basis of a predictive, proscriptive organizationalmodel This can in turn lead to more efficient, robust and context-sensitive systems byincreasing the level of detail at which competing organizational designs are evaluated.

To accomplish this, I introduce a new, domain-independent organizational designrepresentation able to model and predict the quantitative performance characteris-tics of agent organizations This representation, capable of capturing a wide range

of multi-agent characteristics in a single, succinct model, supports the selection of anappropriate design given a particular operational context I demonstrate the repre-sentational capabilities and efficacy of the language by comparing a range of metricspredicted by detailed models of a distributed sensor network and information retrievalsystem to empirical results In addition to their predictive ability, these same mod-els also describe the range of possible organizations in those domains I show howgeneral search techniques can be used to explore this space, using those quantitativepredictions to evaluate alternatives and enable automated organizational design

TABLE OF CONTENTS

Page

ACKNOWLEDGMENTS v

ABSTRACT vii

LIST OF TABLES xiv

LIST OF FIGURES xv

CHAPTER 1 INTRODUCTION 1

1.1 Introduction 1

1.2 Major Ideas 11

1.2.1 Basic Assumptions 11

1.2.2 Organizational Design 13

1.2.3 Representing Organizations 15

1.3 Guide to the Dissertation 17

2 REPRESENTING ORGANIZATIONS 19

2.1 Organizational Effects 19

2.1.1 The Distributed Sensor Network Domain 20

2.1.2 Empirical Demonstration of Organizational Effects 25

2.1.3 Geographic Coalitions 26

2.1.4 Functional Differentiation 27

2.1.5 Organizational Maintenance 29

2.1.6 Generality of Effects 32

2.2 Organizational Representation 34

2.2.1 ODML 36

2.2.2 Distributed Sensor Network Model 46

2.2.2.1 Has-A Relationships 49

2.2.2.2 Is-A Relationships 50

2.2.2.3 Templates and Instances 51

2.2.2.4 Quantitative Expressions 52

2.2.3 Supported Data Types 56

2.2.3.1 Numeric Data 56

2.2.3.2 Lists 57

2.2.3.3 Distributions 57

2.2.4 Limitations and Tradeoffs 62

2.3 Modeling Characteristics in the DSN Domain 65

2.3.1 Environmental and System Constants 65

2.3.2 Local Characteristics 67

2.3.3 Entity Interactions 68

2.3.4 Multiple Role Assignments 70

2.3.5 Dynamic Behaviors 72

2.3.6 Heterogeneity 74

2.3.7 Conflicts, Constraints and Resolution 75

2.3.8 Organizational Utility 78

2.3.9 Evaluation of Representation 81

2.4 Conclusions 84

3 MODELING AN INFORMATION RETRIEVAL SYSTEM 86

3.1 Information Retrieval Model 87

3.2 IR Simulation 91

3.3 Representing IR Characteristics 93

3.3.1 Roles 94

3.3.2 Data Sources and Collection Signatures 94

3.3.3 Probabilistic Search and Query 95

3.3.4 Query Response Time 101

3.3.5 Constraints 113

3.3.6 Organizational Utility 114

3.4 Conclusions 122

4 DESIGNING ORGANIZATIONS 123

4.1 Designing Organizations 124

4.1.1 Design Complexity 127

4.2 Algorithmic Search Techniques 134

4.2.1 Exploiting Hard Constraints 136

4.2.1.1 Monotonic Trends in the DSN Model 141

4.2.1.2 Non-monotonic Trends in a SUBSET-SUM Model 145

4.2.1.3 Results 148

4.2.2 Equivalence Classes 149

4.2.2.1 Results 152

4.2.2.2 Selecting Appropriate Discriminators 153

4.2.3 Using General Mathematical Solvers 155

4.2.4 Avoiding Redundant Search 162

4.2.5 Independent Sub-Problems 163

4.2.6 Domain-Specific Techniques 163

4.3 Heuristic Modeling Techniques 164

4.3.1 Homogeneity 166

4.3.2 Abstraction 167

4.3.3 Heuristic Modeling Results 169

4.4 Designing Organizations in Practice 173

4.4.1 Efficiently Evaluating Organizations 173

4.4.2 Constructing Organizations 177

4.4.2.1 Top-Down Construction 177

4.4.2.2 Bottom-Up Construction 182

4.4.3 Applying Designs to Actual Systems 183

4.4.4 Model Inheritance and Reuse 185

4.5 Adapting Organizations 187

4.5.1 Monitoring for Problems 188

4.5.2 Searching for Solutions 191

4.6 Conclusions 193

5 MODELING OTHER CHARACTERISTICS AND PARADIGMS 194

5.1 Modeling Common Organizational Characteristics 195

5.1.1 Non-Uniformity 195

5.1.2 Temporal Interactions 201

5.1.3 Levels of Abstraction 203

5.2 Modeling Common Organizational Paradigms 205

5.2.1 Hierarchies and Holarchies 205

5.2.2 Coalitions and Congregations 208

5.2.3 Federations 212

5.2.4 Markets 219

5.2.5 Matrix Organizations 222

5.2.6 Societies 225

5.2.7 Teams 228

5.2.8 Compound Organizations 232

5.3 Conclusions 233

6 RELATED WORK 234

6.1 Organizational Representations 234

6.2 Organizational Evaluation and Design 247

7 CONCLUSIONS 256

7.1 Summary 256

7.2 Contributions 262

7.3 Discussion 265

7.4 Future Directions 270

APPENDICES A TRANSLATING ODML TO MATHEMATICA 274

B A SURVEY OF MULTI-AGENT ORGANIZATIONAL PARADIGMS 278

C DISTRIBUTED SENSOR NETWORK ODML MODEL 336

D INFORMATION RETRIEVAL ODML MODEL 346

E TILING REDUCTION MODEL 355

F SUBSET-SUM REDUCTION MODEL 357

BIBLIOGRAPHY 358

LIST OF TABLES

predicted and empirical response recall values from Figure

3.4 100

templates Number of agents and utility are given for the optimalfound organization 170

templates Number of agents and utility are given for the optimalfound organization 172

different organizational representations 247

different organizational design and adaptation schemes 255

A.1 ODML model element to Mathematica translation table 275A.2 ODML built-in function to Mathematica translation table 277

LIST OF FIGURES

network organization 6

2.1 Organization-centric view of the DSN architecture 21

2.2 The DSN architecture in four phases A: sectorization of the environment, B: distribution of the scan schedule, C: negotiation over tracking measurements, D: tracking data fusion 22

2.3 Effect of sector size on messaging 26

2.4 Messaging disparity vs sector size 28

2.5 Message types vs sector size 30

2.6 Effect of sector size on RMS error 30

2.7 Average communication distance 31

2.8 Communication disparity (a) and RMS error differences (b) with varied sector sizes and target densities 33

2.9 Pseudocode for the get value function of a node N This function is used to quantify the characteristics of instance nodes 45

2.10 Example ODML (a) template and (b) instance structures for the sensor network organization 47

2.11 A portion of the ODML specification for the track manager and

s tm relation nodes The structural has-a relations are defined in

track manager first, using an estimate of the role’s migration rate

and the number of extant sensor and sector manager nodes An

estimate of the target bounds is expressed next, and that bound

used with the environmental sensor density to estimate the

number of sensors that will be contacted for data The

sm tm relationshows how modifiers are used to first propagate

that demand to a sensor, and later to inform the track manager of

the the resulting actual measurement rate 48

2.12 An ODML specification for the agent types in the DSN domain

showing how is-a relationships allow characteristics to be shared

among derived nodes, and how those derived nodes can be

differentiated through their additional local definitions 51

2.13 A graph showing some of the equations and equation-based

interactions between nodes that are used to predict RMS trackingerror in the DSN model Solid arrows represent has-a

relationships, while dashed arrows between nodes indicate where amodifier is used to propagate values between nodes 542.14 Pseudocode for the hierarchical Monte Carlo trial procedure 61

2.15 A portion of the ODML specification for the sensor node showing

how message load values, number of sensors controlled and role

indicators are passed to the agent with modifiers 71

2.16 A portion of the ODML specification that estimates the delay in

directory responses that is experienced by a track manager The

sm tm relationnodes determine the individual delays for each

sector manager, while the track manager uses these values to

estimate a worst case delay that is used to reduce the requested

measurement rate 76

2.17 Performance predicted by the ODML sensor network model versus

empirical observations for a) Total messaging, b) Messaging

disparity, c) Message type totals and d) RMS error Predicted

lines are solid, empirical are dashed 80

2.18 A comparison of the average actual and model-predicted

characteristics by role, for agents operating in the distributed

sensor network 82

2.19 Performance predicted by the ODML model as the number of targets

is varied for a) Messaging disparity, and b) RMS error 83

query in the information retrieval organization Steps in the

numbered trace correspond to messaging events, except for

bracketed elements that indicate local processing 88

domain b) A small organizational instance produced from that

template 90

across a range of search and query size parameters 99

nodes detailing the fields used to estimate response time 105

distribution 106

sets of databases 107

ODML model and observed in organizations with (a,b) five

[1M,4D], (c,d) fifteen [1M,6A,8D], and (e,f) twenty-eight

[1M,7A,20D] agents In the designs shown in (a,c,e), node M is

the mediator, A are aggregators, and D are databases 112

organizations when the query rate (queries per second) is varied

Mediators and aggregators are shown as hollow circles, while the

solid databases form the leaves Higher is better, optimal values

for each rate are shown in bold 116

3.10 Similar to Figure 3.9, the utility predicted for the range of possible

six-database organizations when the query rate (queries per

second) is varied, but with the query and search sizes set to one

Note differences in organizations 11-18 119

3.11 The range of optimal organization instances derived from the IR

template, when 1,3,6,9,12 or 15 databases are available The

environment and organization nodes have been omitted for

clarity 120

3.12 A comparison of the waiting time distributions for the different

optimal organizations from Figure 3.11 The distributions widen

and shift right as the number of databases increase 121

problem (bottom) A valid organizational instance created from

constraint estimation algorithm across three different models 147

and without equivalence classes across four differently sized sensornetworks The models allowed designs with 3, 6, 9 and 12 total

sensors, respectively 152

engine The original template (a), the instance with deferred

variables (b), and the final organizational instance (c) 159

4.10 The result of the ODML translation process This code is passed into

Mathematica to perform the optimization 1604.11 Performance of the numeric solver versus conventional search across

different problem complexities 161

4.12 Two information retrieval templates, derived from Figure 3.2a a)

Incorporates homogeneity, by limiting aggregator selection to twodistinct choices b) Incorporates abstraction, by eliminating the

4.13 The optimal organization found by the baseline template for the

small-scale scenario 170

4.14 Optimal organization instances produced by the Homogeneous +

Abstract (a) and (b) templates for the small-scale scenario 1714.15 Search performance with and without cached values 175

4.16 The partitioning of a local search tree Strikeouts indicate visited

choices, bolded are the current choice, while a question mark

indicates additional unknown choices remain 179

4.17 Characteristics of the distributed search using different numbers of

processing agents The speedup factor is shown in (a), and the

number of messages required in (b) 1804.18 Reusing common agent models in new domains 186

from the original DSN model are shown in the shadowed

boxes 196

model This reflects the sensor’s range of approximately 30

feet 197

model using non-uniform target paths (upper) and sensor

locations (lower) The layouts (a,c) show the sensor and target

arrangements for two different scenarios Sensors are circles,

targets are triangles and their paths are dashed lines The graphs(b,d) show the demand levels by sensor (top, darker is target 0,

lighter is target 1) and the measurement levels by track manager

(bottom, darker is the amount received, lighter is the additional

amount that was requested but not able to be satisfied) 199

the effects that shifting starting points have on performance

Legend is the same as Figure 5.3 203

5.5 A range of modeling possibilities, each with different levels of

abstraction 204

5.6 Using a hierarchy template in the IR domain The design shown in Figure 3.2b shows a possible corresponding instance 206

5.7 An ODML holarchy a) template and b) example instance The structure is quite similar to the hierarchy in Figure B.1, differences lie primarily between their internal behaviors 207

5.8 An ODML coalition a) template and b) example instance 208

5.9 An ODML federation a) template and b) example instance A federation consist of a set of members that have relinquished control to a distinguished intermediary The broker and reflector nodes represent two possible intermediaries Only intermediaries interact between federations, as represented by the links through relationship nodes in b) 212

5.10 An alternate ODML federation template that allows greater differentiation among intermediary styles but is less able to inherit and reuse common elements 217

5.11 An ODML market a) template and b) example instance 219

5.12 An ODML matrix a) template and b) example instance The wm-relation node is used to propagate the effects of each manager to a worker 223

5.13 An ODML team a) template and b) example instance Has-a relationships from teams to goals are used to allocate work among the entities 229

5.14 Portions of an ODML team model supporting goal alternatives 231

B.1 A hierarchical organization 280

B.2 A holarchical organization 285

B.3 A coalition-based organization 289

B.4 A team-based organization 293

B.5 Congregations of agents 299

B.6 An agent society 303B.7 An agent federation 309B.8 A multi-agent marketplace 313B.9 A matrix organization 319B.10 A compound organization 322B.11 Comparing the qualities of various organization paradigms 331

C.1 Graphical view of the ODML DNS model 337

D.1 Graphical view of the ODML information retrieval model 347D.2 Graphical view of the complete ODML information retrieval model

obtained from the source in this appendix 348

CHAPTER 1 INTRODUCTION

1.1 Introduction

Many of the decisions made in multi-agent system design, and in computationalsystems in general, are predicated on the idea that one wishes to minimize the “bad”characteristics of the system while maximizing the “good” This practice manifestsitself in blanket, axiomatic objectives such as “minimizing communication”, “reduc-ing uncertainty”, and “maximizing profit” While these are worthy, abstract goalsthat have critical practical and research importance, when a system is deployed andsituated in context, such ideals may no longer have the same level of relevance Con-sider the underlying issues that drive these objectives Why should communication

be minimized? Why do we care about the combinatorics of a particular technique?Why should centralization be avoided? In each case, we presume the existence ofsome limiting factor, some bounded resource that motivates these objectives How-ever, when the system is placed in a particular context where these bounds can bequantified, the intangible nature of these blanket statements is no longer sufficient.For example, if ample communication bandwidth is available and additional utilitymay be derived by using it, then a strategy that always minimizes communicationmay lead to a solution that fails to reach its potential If a particular resource isbounded, and the qualitative side effects of using some or all of that resource arethe same, then the system should exploit it as best it can in service of satisfying ormaximizing the system’s specified goals

Because of this, I believe that any real-world system must be tailored to the ronment in which it exists, if it is to make effective use of the resources and flexibility

envi-available to it I explore this tailoring through the system’s organizational design.

The notion of an organizational design is used in many different fields, and generallyrefers to how members of a society act and relate with one another This is true

of multi-agent systems, where the organizational design of a system can include adescription of what types of agents exist in the environment, what roles they take on,and specifications guiding how they act both independently and with one another.More generally, if we assume an entity has a set of possible choices to make dur-ing its operation, the organizational design will identify a particular subset of thosechoices that should actually be considered at runtime By working with this typi-cally smaller set, the entity’s decision process is facilitated This additional structurebecomes increasingly important as the system scales in number and scope [32] Imag-ine how difficult it would be for a large human organization, such as a corporation

or government, to function if individuals lacked job descriptions and long-term peerrelationships Agent systems face similar challenges, and can derive similar benefitsfrom an explicit organizational design

Consider the problem of designing a solution for a complex, resource-boundeddomain, such as a distributed network of sensors that is used for tracking Suchsystems typically consist of an array of sensor nodes that are deployed to obtain themeasurement data needed to track mobile targets in an environment Assume in this

case that each sensor is host to a local process called an agent that is responsible for

controlling the sensor Let us further assume that the sensor nodes must collaborate

in some way to be successful, because multiple sensors must illuminate a target multaneously to correctly obtain its position Given these assumptions, a designermust determine a way to structure the agents’ behaviors so that tracking may be ac-

si-complished One strategy would create or delegate a single agent to be the manager

of the entire sensor network The manager would decide when, where and how eachsensor should take measurements, and then process the resulting data to estimate thetargets’ positions This layout of responsibilities constitutes a rudimentary organi-zational design It specifies what roles agents take on, who they interact with, andwhere decision making authority is located.

Under some conditions, this simple solution will perform optimally, because themanager can maintain an omniscient view of the entire network’s state and use thatview to find the best assignment of sensing tasks However, under real world condi-tions, where bandwidth and computational power is limited, communication and dataprocessing takes time, and the number of sensors can be arbitrarily large, the weak-nesses of this approach quickly become apparent A different strategy, in the form of adifferent organizational design, can compensate for these more challenging conditions.For example, we might distribute the manager role among multiple agents to moreevenly balance the communication and computational loads We might also create aninformation dissemination hierarchy among the agents that prioritizes, summarizes,and propagates measurement data to use the available bandwidth more efficiently.However, distributing the role can lead to conflicts among managers and lower util-ity assignments, because no single agent necessarily has the local context to makethe right decision Similarly, the summarization process of a hierarchical distributionscheme can introduce additional latency and imprecision Because of these tradeoffs,the organization can be a double-edged sword, both helping and hindering the system

in potentially complex ways The questions I address in this thesis revolve aroundfinding a general way to determine the most appropriate organizational strategy for

a given situation when there are many such strategies to consider

Implicit in this example is the idea that different organizations will affect theperformance of a working system in different ways Intuitively, changing the manner

in which agents interact or the pattern that those interactions take on can change

how the system behaves from both global and local perspectives The objectives of aparticular design will depend on the desired solution characteristics, so for differentproblems one might specify organizations which aim toward scalability, reliability,speed, or efficiency, among other things Confounding the search for such a design

is the fact that many potentially important characteristics can be subtle, not readilyidentified as the system is being developed, or have complex interactions

For example, at what point do the benefits of the dissemination hierarchy proposedabove outweigh its costs? The additional communication and processing resourcesrequired to implement it may not be readily available Obtaining them may require

a monetary cost if new systems must be purchased, or a complexity cost if the newresponsibilities are spread among the existing systems At the same time, one mustreason about the dimensions of the hierarchy – how tall and wide should it be? Whichentities should be assigned the responsibilities present at each node? Should the treedimensions be kept small, potentially concentrating the burden, or be made large tomore evenly distribute the load? The designer will likely have an intuitive grasp ofwhat is required, which is how existing systems are typically developed However,all of these features are interrelated along with the goals of the system, the expectedtasks it will experience, and the nature of the available resources Intuition can fallshort when such interactions allow small changes to lead to unexpected outcomes.These so-called phase transitions or tipping points require a deeper understandingand a more concrete representation to be addressed

Although individuals have created mathematical models for particular aspects oforganizationally-driven agent behaviors [166, 123, 163, 39, 68, 160], none have ex-plored the utility of a general modeling language capable of incorporating arbitraryquantitative information It is my belief that understanding the fundamental causes

of characteristics like those described above, and using that information to developaccurate, predictive models of their effects are both critical to selecting an appropriate

design, particularly as the agent population grows in scale or complexity If we are

to understand these effects and develop the means by which they can be exploited oravoided through organizational design, we must have a representation capable of ex-pressing the range of ways the design can be created and capturing the characteristicseach design will exhibit

Many different representations have been created to describe agent organizations[187, 40, 109, 143, 45, 171, 50, 59, 122, 88, 174] Most fall into one of two categories:either they represent a wide range of organizational characteristics abstractly, or theycan capture a smaller set of characteristics concretely The former are usually good atrepresenting what entities exist or could exist, but cannot compare alternatives in aquantitative way The latter may contain quantitative knowledge, but have difficultyrelating that knowledge to specific organizational concepts, mitigating their usefulness

if one is hoping to understand the effects a particular organizational design will have.More specifically, existing organizational representations are either flexible andqualitative or inflexible and quantitative In this work I demonstrate that it is possible

to create a representation that is both flexible and quantitative I introduce a newrepresentation, the Organizational Design Modeling Language (ODML), designed tocapture organizational information in a single unified, predictive structure Thisrepresentation, described in detail in Chapter 2 has the capability to model a widerange of organizational paradigms and characteristics, at different levels of abstractionacross many different domains At the same time, it is able to integrate concretenumeric information in the form of expressions and predictive equations using a range

of mathematical techniques Using this representation, it is possible to create a widerange of integrated models that possess a level of quantitative detail that is notpossible with existing languages

ODML models are superficially graph-based structures that consist of nodes andedges Figure 1.1a shows an example ODML template model from a more detailed

Of equal importance are the quantitative details that exist within each node, whichare not shown in Figure 1.1 Each node contains a set of fields that quantitatively de-

scribe the relevant characteristics of the node using mathematical expressions Theseexpressions can affect or be affected by the characteristics of other nodes, implicitlycreating a second, more detailed layer of relationships through which one aspect ofthe organization may affect another In this way, quantitative information about theenvironment, resources, agents, tasks, goals, or other components relevant to the sys-tem’s performance can be incorporated into a single model and tied together in a waythat captures the interdependencies that exist between them This web of equations

is a key component of ODML that differentiates it from existing representations

As mentioned above, understanding the quantitative effects of organization is

a necessary prelude to determining which design is most appropriate for a givenoperational context ODML’s capacity to represent detailed numeric informationtherefore gives it part of the functionality needed to address this problem To providethe remaining requisite functionality, it must also be able to express the variety of waysthe design can be manifested This is accomplished in ODML through the creation

of a template (Figure 1.1a) that has decision points embedded within it Differentchoices for those decisions will result in different candidate designs It is in this waythat the template embodies the space of organizational alternatives Given a templateand a set of choices for those decisions points, one can create a particular instance ofthat organization (Figure 1.1b) In that example two sectors have been created, eachwith a sector manager (SM) and two sensors (S) A single track manager (TM) isconnected to these entities through a series of relationships By using the embeddedexpressions in the instance model to relate, predict and evaluate its characteristics,the instance may be automatically compared and ranked against other competingdesigns This ranking is then used as part of a search process to select the mostappropriate design

It is during this evaluation and ranking phase that the web of equations is usedmost Other representations typically perform their evaluation using a fixed set of

characteristics limited by the language, through simulations or model-specific tic analyses, or through more qualitative or logical comparisons ODML is differen-tiated by the fact that one can embed arbitrary mathematical expressions within themodel, and use those to produce fast, precise predictions of whatever characteristicsare deemed relevant to evaluating design utility.

heuris-The flexibility of the ODML representation, its ability to model a wide range ofconcepts and functionality, is derived from the nature of the language itself Nearlyall existing organizational representations are structured around a well-defined set ofrequired or permissible structures For example, they will have concrete and explicitnotions of an agent, a role, norms or goals These concepts can be represented inODML, but this representation is accomplished using only the primitive notions ofnode, relationship and quantitative characteristics outlined above; they have no pre-defined semantics

Although having such built-in structures can be beneficial, their existence, ularly if they are required, means that any model created with such a language mustabide by the assumptions associated with those structures These assumptions can

partic-be sufficiently constraining or inflexible that the representation is no longer usable,

or that the accuracy of the resulting model is compromised For example, a proposedorganization may be large enough that one would not want to have an explicit no-tion each individual agent, because to do so would result in a model so large as to

be impractical In other cases the agent characteristics supported by the languagemay be insufficient or inappropriate to capture the nature of the domain in question.Because ODML makes no such assumptions, the first designer could choose to omit

an agent node, creating a more abstract but correspondingly more scalable model.The second designer has the freedom to incorporate only those characteristics theydeem appropriate ODML’s relatively primitive language structure is another char-acteristic that differentiates it I will demonstrate how the flexibility this primitive

nature provides allows it to model a range of detailed concepts not easily captured inother representations.

The drawback to having a language lacking in high-level concepts is that it makesthe search process more difficult For example, in languages that explicitly supportagents and roles it is possible to embed into the search process the idea that roles must

be bound to agents, and create heuristics or strategies designed to take advantage ofthat special relationship (e.g., [174, 159, 138, 35]) Lacking such structural landmarks,

a search process can only attempt to infer that this is the case Given that such arelationship might be modeled in different ways by different designers (or omittedaltogether as above), the applicability of such concept-specific techniques is limited.Because of this, and because the search space created by an ODML template can

be quite large, solving the search problem is difficult In Chapter 4 I demonstrate thisdifficulty by proving that the problem of finding a valid instance within an ODMLtemplate space is NEXP-complete That chapter also describes a range of techniquesthat can be used to exploit the mathematical nature of the representation and makethe search more efficient In particular, I show it is possible to exploit constraintsembedded in the model to bound the search space, and to devise notions of choiceequivalence to avoid redundant parts of the space Both techniques can result in asignificant reduction of the time required to search I will also demonstrate techniquesthat change the model itself, by incorporating abstraction or homogeneity to reducethe search space Finally, a distributed search process is described that attacks thecomplexity through parallelism

To evaluate the efficacy of ODML and the quality of the automated design process,

I have developed complete models for two different operational domains: a distributedsensor network and a peer-to-peer, distributed information retrieval system Thedistributed sensor network model has been created based on the design of an existing,real-world architecture [86] The original system was developed and analyzed prior to

the creation of ODML, and thus provides a unique opportunity to evaluate ODML’sability to represent and predict a number of organizational characteristics Thisenvironment and model are described in detail in Chapter 2 The second model isinspired by an information retrieval system developed by Zhang and Lesser [215] Itfeatures a network of flexible hierarchies that can be structured in a vast number ofways, providing a space of designs that is large enough to make simple brute-forceapproaches impractical The interactions between entities are also complex, requiringthe integration of different mathematical techniques to be correctly captured Thiswork is described in Chapter 3 The predictions produced by both models are alsoempirically verified in those sections Several other models are also presented inChapter 5 as a demonstration of its applicability.

To summarize, the primary objective of this dissertation is to demonstrate thefeasibility of a highly quantitative representation and the increased utility that such

a representation brings to the organizational design problem The following butions will be made to that end

contri-1 I show that a flexible, accurate organizational representation grounded in bitrary quantitative information can be created This is accomplished throughthe design and implementation of the ODML language itself

ar-2 I show that it is possible to use such a language to quantitatively model andaccurately predict the complex, interrelated characteristics of organizations op-erating in realistic domains This is demonstrated through the creation andempirical evaluation of several ODML models that address different problems

in different domains, each of which has a different set of relevant characteristics

3 I demonstrate that such models are capable of capturing a range of differentorganizational designs at different levels of abstraction This is accomplished byintegrating decision points within ODML models that reflect the choices that

must be made to create an organizational instance, and by demonstrating that

by varying such choices different designs will be produced

4 I demonstrate that techniques can be devised that automatically search andevaluate the design space to solve the organizational design problem This isaccomplished by devising techniques that use the mathematical substructure tobound the search space, and use quantitative predictions of instance behaviors

to evaluate and rank competing designs

I will return to this list of contributions in Chapter 7 to provide additional details

on how they have been accomplished

1.2 Major Ideas

This section continues introducing the dissertation by enumerating the basic sumptions I have made when pursuing the objectives outlined above Sections 1.2.2and 1.2.3 expand upon the high-level concepts that support and motivate this workthrough additional discussion of the organizational design and representation prob-lem

The strategy that I present to satisfy the objectives outlined above assumes that

it is both possible to model the system in question and that different systems havemeasurable differences More completely, I assume the following conditions are true:

1 The characteristics of the environment, resources, agents, tasks, goals, or anyother object relevant to the system’s performance are known, and they may beeither determined exactly or approximated This may be accomplished ana-lytically, through repeated observation of the characteristic in question, or bedefined by some other external process

2 There are a range of design decisions that can be made, resulting in an array ofcandidate designs that exhibit different characteristics.

3 Active environmental entities, such as agents, are able to appropriately porate and respect any design choices specified by the selected model

incor-4 Quantifiable, measurable characteristics exist that can be used to differentiatecandidate systems

5 The set of characteristics relevant to the system’s intent can be combined insome manner to produce a single, numeric value that can be used to assign a

preference order to the candidates This value is the system’s utility.

With the possible exception of point 1, these same assumptions are also made

in the majority of related work The essence of point 1 is an assumption that theunderlying mechanics of the system are either known or can be determined throughanalysis This does not mean that there is no uncertainty, but it does mean that thelevel of uncertainty is also known, so that it can be represented and reasoned about

I will argue in Section 4.5 that, in the absence of perfect data, runtime adaptationcan be still used to address problems that arise from designs based on incorrect orout-of-date knowledge

I do not make any strong assumptions about the underlying architecture employed

by agents or other entities within the organization, except to the extent that: 1) theycan correctly enact relevant organizational decisions suggested by the model and2) they exhibit behaviors that can be captured by that same model Section 2.2demonstrates that even complex inter- and intra-agent behaviors can frequently becaptured with sufficient fidelity by a succinct set of expressions [84] By employingabstraction or statistical characterizations, the germane aspects of such phenomenacan be represented, reasoned over, and guided by an organizational design

Later sections will present organizational models where additional assumptionsmay be made concerning the behaviors, conditions, or other features of the specificagents, environment or resources depicted in those models I describe any such as-sumptions within their respective contexts.

The organizational design of a multi-agent system is the collection of roles, tionships, and authority structures that govern the system’s behavior All multi-agentsystems possess some or all of these characteristics, and therefore all have some form

rela-of organization, although it may be implicit and informal Just as with human nizations, agent organizations guide how the members of the population interact withone another, not necessarily on a moment-by-moment basis, but over the potentiallylong-term course of a particular goal or set of goals This guidance might influencethe data flow, resource allocation, coordination patterns or any number of other sys-tem characteristics [74, 23] This can help groups of simple agents exhibit complexbehaviors, and help sophisticated agents reduce the complexity of their reasoning.Implicit in this concept is the assumption that the organization serves some pur-pose — that the shape, size and characteristics of the organizational structure caneffect the behavior of the system [62] It has been repeatedly shown that the organi-zation of a system can have significant impact on its short and long-term performance[24, 156, 84, 128, 6, 180, 17], dependent on the characteristics of the agent population,scenario goals and surrounding environment Because of this, the study of organi-zational characteristics, generally known as computational organization theory, hasreceived much attention by multi-agent researchers

orga-An organizational design can influence the system at many different levels ofabstraction For example, consider the sensor network variants discussed in Section1.1 One aspect of the design might dictate how the agents are arranged from a high-

level perspective, in this case using a completely centralized or a more distributedtopology This aspect is typically associated with deciding or putting limits on entityinteractions For example, in the centralized configuration, all sensors directly interactwith the manager but never with each other A design can also influence internalbehaviors, such as specifying managerial or authority relationship For example, themanager was given the authority to assign tasks to individual sensors in the network.

I also believe that an organizational design’s influence can extend to almost anyaspect of any entity in the environment, agent or otherwise, that is relevant to thesystem’s performance; although this expanded definition is not universally held Forexample, I assume an organizational design can specify the protocols that agents use

to communicate, the algorithms used to find peers, and even the manner in whichagents resolve purely internal conflicts This is because these low-level characteristicscan affect the performance of the conventionally high-level organizational structureand vice versa Therefore, when selecting an appropriate design, all relevant factorsshould be considered, regardless of where in the architecture they may occur Thisview is consistent with ODML’s approach to specifying design alternatives, which per-mits design alternatives to be modeled as an internal agent decision while permittingthe ramifications of that decision to propagate elsewhere in the model

It is generally agreed that there is no single type of organization that is suitablefor all situations [91, 32, 112, 24] In some cases, no single organizational style is ap-propriate for a particular situation, and a number of different, concurrently operatingorganizational structures are needed [63, 86] Some researchers go so far as to say noperfect organization exists for any situation, due to the inevitable tradeoffs that must

be made and the uncertainty, lack of global coherence and dynamism present in anyrealistic population [150] Although I do not subscribe to this particular view, what

is agreed upon is that all approaches have different characteristics that may be moresuitable for some problems and less suitable for others

Organizations can be used to limit the scope of interactions, provide strength innumbers, reduce or manage uncertainty, reduce or explicitly increase redundancy orformalize high-level goals that no single agent may be aware of [113, 61] At thesame time, organizations can also adversely affect computational or communicationoverhead, reduce overall flexibility or reactivity, and add an additional layer of com-plexity to the system [84] By discovering and evaluating these characteristics, andthen encoding them using an explicit representation, one can facilitate the process

of organizational self-design [31] whereby a system automates the process of

select-ing and adaptselect-ing an appropriate organization dynamically [112, 161] This approachwill ultimately enable suitably equipped agent populations to organize themselvesautonomously, eliminating at least some of the need to exhaustively determine allpossible runtime conditions a priori Before this can occur, the space of organiza-tional options must be mapped, and their relative benefits and costs understood

In the preceding section I repeatedly implied that there is or must be some way

of explicitly representing the organization For example, if one is to understand theorganization’s characteristics in a concrete and manipulable sense, there must besome data structure capturing those characteristics Similarly, if one is attempting

to predict the characteristics of a theoretical organization, there must be some modelcapturing the relevant features that underlie those characteristics

Not all applications demand the creation of an explicit organizational design Infact, most agent and distributed systems function without one In these systems theorganization still exists, but it is generally defined only implicitly The roles enti-ties take on may be homogeneous or hard-coded Relationships may emerge throughsearch or happenstance Local behaviors are hard-wired or heuristically driven Nosingle, coherent model ties everything together Clearly these systems perform ade-

quately, or they would have little to demonstrate I do not believe that an explicitrepresentation is essential, but I do contend that it can be eminently useful For exam-ple, a deep understanding of the ramifications of role assignment requires some model

of what the role demands of the agent it is assigned to, and how those demands relate

to other aspects of the agent’s existence Knowing when and with whom to form arelationship can be decided locally, but the more global consequences of this choicecan be made more apparent by recognizing the potentially nonlocal effects it canhave The same is true of local decision making patterns – discovering and evaluatingthe nonlocal effects of these decisions can be greatly facilitated by an appropriate,explicit organizational representation

The computational structure one uses to formally encode the organization’s acteristics is the foundation upon which other organizational activities are based.Thus, if the structure is lacking in fidelity or capability, some activities may provedifficult or impossible to accomplish For example, if the organization does not con-tain quantifiable measures of load or performance, it will be difficult for the agents

char-to discriminate among competing strategies based on that criteria Similarly, if thestructure is too abstract, it may be unable to support the practical decisions thatmust be made during an agent’s lifetime Conversely, if it is too complex, it mayprove to be computationally infeasible to work with the structure Thus, there is atension between expressiveness and ease of use that must be addressed when select-ing or defining an organization description language ODML’s approach is in generalmore complex and detailed than related representations, resulting at the same time inmore precise information and a more difficult search process I believe this tradeoff iswarranted by the additional precision it can produce, leading to more robust designsand a model that has value in a larger number of situations

1.3 Guide to the Dissertation

Chapter 2 describes the organizational representation problem It begins in tion 2.1 with an example of an organizationally-driven solution to the distributedsensor network problem mentioned above A corresponding set of experiments pro-vides concrete examples of how the design of an organization can affect performance,which motivates this work Section 2.2 continues by introducing ODML, providing aformal description of the structure and built-in functionality A detailed example ofits use modeling the distributed sensor network system is given, along with empiricaltests validating the predictions made by the model

Sec-Chapter 3 describes and validates a detailed model of the information retrievaldomain This is used as a further example of ODML’s flexible yet precise nature

In particular, the mathematics involved in correctly modeling the system are moresophisticated than those used in the distributed sensor network domain Section 3.3details these formulations, and shows how the predictions they make can be used tofind the optimal organization as environmental conditions change The search space

in this domain is large, which motivates the development of search techniques able toexploit the underlying structure

Chapter 4 demonstrates how ODML can be used as the basis for designing nizations, using the DSN and IR models as examples The complexity of the validinstance search is proved in Section 4.1.1 Section 4.1 describes the design search pro-cess itself, and details the algorithmic and modeling techniques that I have created toassist with that search The chapter concludes with Section 4.5, which outlines howODML can be used to address the online adaptation of organizations

orga-Chapter 5 provides additional examples of ODML’s use across a range of featuresand organizational paradigms In particular, Section 5.1 shows how the distributedsensor network model from Chapter 2 can be enhanced to model geographic hetero-geneity, temporal interactions and different levels of abstraction Section 5.2 describes

how models can be created to capture aspects of several additional organizationalparadigms.

Although a brief review of related approaches to these problems is provided alongwith ODML’s introduction in Section 2.2.1, a more thorough description is delayeduntil Chapter 6 This chapter describes related representation and design researchand contrasts those projects with ODML’s capabilities Chapter 7 concludes with asummary of the work, and describes the contributions and conclusions that have beenmade

There are several appendices that provide additional detail and context Appendix

A describes how parts of ODML model and search space can be translated for use

in a general mathematical solver Appendix B is a survey of common organizationalparadigms used in multi-agent systems, which is used as the basis for the discussion inSection 5.2 The remaining Appendices C, D, E and F contain the complete textualsource descriptions of several of the ODML models described in the dissertation